This study proposes an interpretable Graph Neural Network (GNN) framework for rapid linear-elastic seismic displacement prediction of reinforced concrete (RC) frames. Both Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) were validated on 2400 datasets comprising 12 RC frames (3–7 stories) and 200 artificial ground motions, achieving high accuracy (R 2 =0.920 ± 0.012 and R 2 =0.910 ± 0.010, respectively) with no statistically significant difference (p = 0.215). GAT's attention mechanism provides physical interpretability: columns received 50.58 ± 2.55% higher attention than beams (p < 0.001), consistent with vertical load transfer mechanics. Attention also varied by floor level, with ground-level edges receiving highest attention, correlating with lower prediction errors at boundary-constrained locations. Feature ablation on the GAT model identified spatial coordinates as the most critical input (removing them reduces R² by 11.0%, p < 0.001); symmetric ablation further showed that the redundant stiffness features act as training noise, with their removal improving R² by 5.9% (p < 0.001). The proposed framework offers a robust and interpretable tool for seismic assessment of RC frame structures.
Lee et al. (Sat,) studied this question.